Artificial Intelligence

OpenAI Says Its Models Can Help Banks Migrate From Old COBOL

OpenAI states that its latest models feature enhanced capabilities for handling COBOL and other legacy programming languages, offering potential relief to financial institutions facing a workforce shortage.

Radwan AbdulhadiEditor and technology writer

Published: October 20, 2026 · Updated: October 20, 2026 · 3 min read

OpenAI Says Its Models Can Help Banks Migrate From Old COBOL — Image: Engadget

Addressing the Legacy Code Challenge

OpenAI announced that its models now come with improved capabilities for working with COBOL and other older programming languages. According to the company, models are now better at building and migrating software in legacy languages that suffer from a dwindling number of programmers, including Fortran and Visual Basic.

As the company explains in its official documentation, 40 percent of the world's banks still rely heavily on COBOL systems. Migrating these core systems to newer languages would require massive investments costing millions to hundreds of millions of dollars, alongside significant amounts of time.

However, institutions face mounting pressure to upgrade because programmers who work with COBOL are aging out of the workforce. Consequently, banks utilizing these legacy systems will need to execute migrations within the next few years.

Model Evaluations and MultiLangBench

OpenAI noted that its work on improving these legacy language capabilities was guided by direct feedback from enterprise customers and developers. Coding models have traditionally struggled when tasked with handling older languages.

To measure performance, OpenAI created an AI evaluation suite known as MultiLangBench. This benchmark tasks models with creating SQLite-compatible database engines in specific programming languages to test their proficiency.

According to the evaluation data, GPT-6 Astra achieved an average test pass rate of 87.4 percent across 20 languages, showing a significant performance leap compared to GPT-5.6 Sol, which logged a 26.8 percent pass rate. Despite these gains, COBOL remains the most difficult language to crack, yielding a 72 percent pass rate for Astra.

Behavior-Preserving Migrations

To test real-world migration viability, OpenAI evaluated how well its models could translate COBOL programs into modern alternatives like C++, Java, Kotlin, Python, or Rust while strictly preserving the original functionality.

For AI tools to successfully assist financial institutions with modernization, they must accurately preserve specific behaviors of legacy systems. This includes precise technical details such as how numbers with excess decimal digits are rounded and how missing records are handled during execution.

In this specific evaluation category, GPT-6.1 Sol delivered the highest accuracy rate at 97.9 percent when operating with high reasoning effort. When configured to low reasoning effort, the same model achieved an 86.6 percent accuracy rate.

Production Deployment and Necessary Precautions

In its published findings, OpenAI acknowledged that a software migration can only be considered truly successful if all underlying tests pass. Critical infrastructure programs handling sensitive banking operations demand exceptional levels of accuracy.

The company emphasized that behavior-preserving capabilities offer valuable assistance to technical teams maintaining and modernizing older systems. Nevertheless, applying these outputs directly to production software still requires thorough expert review and rigorous testing against established business rules and system dependencies.